Provides simple toy datasets to test on common deep neural network (DNN) architectures and perform layerwise relevance propagation using iNNvestigate using the provided example notebooks. Notebooks also explore the impact of adding informative variables until the network has an AUC of the ROC of 1.0, and a loss of ~1 (exhausting the phase space of the network).
| File | Purpose |
|---|---|
| makeJetImages.ipynb | Makes toy "jet" images (based on Jet-Images -- Deep Learning Edition. |
| fourvec_data.ipynb | Makes toy "four vector" data in the format (pt, eta, phi, mass). |
| CNN_2D.ipynb | 2D model that classifies the images as "signal" or "background" is defined and trained. |
| LRP_CNN2D.ipynb | Takes 2D CNN and jet images as input and creates heat maps of the relevant input features. |
| CNN_1D.ipynb | 2D model that classifies the toy vectors as "signal" or "background" is defined and trained. |
| MLP_multin.ipynb | Feed-forward dense network with only the variables used to make the jet images, and not the images themselves as inputs. |
| deprecated | Holds older versions of CNN's and data making scripts. |